Characterize the future task information encoded by human gaze

Characterize the broader future task information encoded by human gaze in long-horizon and multi-object manipulation tasks, where analogous predictive baselines are less straightforward than the kinematic predictors evaluated for dynamic interception.

Background

The paper compares predicted human gaze with current-position detection, Kalman-extrapolated future position, random crosshairs, and no cue on the Catch Rolling Ball task. The results indicate that predictive position information improves interception performance and that human gaze can anticipate upcoming task-relevant locations.

For long-horizon or multi-object manipulation, however, constructing comparable predictive baselines is more difficult because the relevant future object or placement target may depend on task phase, object identity, and extended action sequences. The paper therefore leaves unresolved how to broadly characterize the future task information conveyed by gaze in such settings.